T

Surface Roughness Prediction for CNC Milling

Process using Artificial Neural Network

AbstractIn CNC milling process, proper setting of cutting parameter is
important to obtain better surface roughness. Unfortunately, conventional try
and error method is time consuming as well as high cost. The purpose for this
research is to develop mathematical model using multiple regression and artificial
neural network model for artificial intelligent method. Spindle speed, feed rate, and
depth of cut have been chosen as predictors in order to predict surface
roughness. 27 samples were run by using FANUC CNC Milling -T14E. The
experiment is executed by using full- factorial design. Analysis of variances shows
that the most significant parameter is feed rate followed by spindle speed and lastly
depth of cut. After the predicted surface roughness has been obtained by using both
methods, average percentage error is calculated. The mathematical model
developed by using multiple regression method shows the accuracy of 86.7% which
is reliable to be used in surface roughness prediction. On the other hand, artificial
neural network technique shows the accuracy of 93.58% which is feasible and
applicable in prediction of surface roughness. The result from this research is
useful to be implemented in industry to reduce time and cost in surface roughness
prediction.
KeywordsCNC milling, surface roughness, multiple regression, artificial
neural network
I. INTRODUCTION
HE challenge of modern machining industries is mainly focused on the achievement of high
quality, in term of work piece dimensional accuracy, surface finish, high production rate,
less wear on the cutting tools, economy of machining in terms of cost saving and increase
of the performance of the product with reduced environmental impact. End milling is a
very commonly used machining process in industry. The ability to control the process for better
quality of the final product is paramount importance.
The mechanism behind the formation of surface roughness in CNC milling process is very
dynamic, complicated, and process dependent. Several factors will influence the final surface
roughness in a CNC milling operations such as controllable factors (spindle speed, feed rate
and depth of cut) and uncontrollable factors (tool geometry and material properties of both
tool and workpiece). Some of the machine operator using trial and error method to set-up
milling machine cutting conditions (Julie Z.Zhang et al., 2006). This
method is not effective and efficient and the achievement of a desirable value is a repetitive and empirical process that can be very
time consuming.
Thus, a mathematical model using statistical method provides a better solution. Multiple regression analysis is suitable
to find the best combination of independent variables which is spindle speed, feed rate, and the depth of cut in order to achieve
desired surface roughness. Unfortunately, multiple regression model is obtained from a statistical analysis which is have to collect
large sample of data. Realizing that matter, Artificial Neural Network (ANN) is state of the art artificial intelligent method that
has possibility to enhance the prediction of surface roughness.
This paper will present the application of ANN to predict surface roughness for CNC milling process. The accuracy of ANN
to predict surface roughness will be compared with mathematical model that built using multiple regression analysis.
II. LITERATURE REVIEW
Previously, Oktem et al. (2005) proposed the genetic programming approach to predict surface roughness based on cutting
parameters (spindle speed, feed rate and depth of cute) and on vibrations between cutting tool and workpiece. From this research,
they conclude that the models that involve three cutting parameters and also vibrating, give the most accurate predictions of
surface roughness by using genetic programming.
Later on 2007, Chang et al. were established a method to predict surface roughness in-process. In their research,
roughness of machined surface was assumed to be generated by the relative motion between tool and workpiece and the
geometric factors of a tool. The relative motion caused by the machining process could be measured in process using a
cylindrical capacitive displacement sensor (CCDS). The CCDS was installed at the quill of a spindle and the sensing was
not disturbed by the cutting. A simple linear regression model was developed to predict surface roughness using the measured
signals of relative motion. Surface roughness was predicted from the displacement signal of spindle motion. The linear regression
model was proposed and its effectiveness was verified from cutting tests.
Tasdemir et al. (2008) applied ANN to predict surface roughness a turning process. This method was found to be quite
effective and utilizes fewer training and testing data.
1i 2 1i 2i 3 1i 3i 1i i

3
Hazim et.al (2009) developed a surface roughness model in
end milling by using Swarm Intelligence. From the studies,
data was collected from CNC cutting experiments using
Design of Experiments approach. The inputs to the model
consist of Feed, Speed and Depth of cut while the output from
the model is surface roughness. The model is validated
through a comparison of the experimental values with their
predicted counterparts.
III. METHODOLOGY
m
(6)
A. Multiple Regression Analysis
x
i
1

ji
u
i
+

j
j = 1 to n (7)
After the surface roughness is obtained for all experiments,
and between hidden layer and output layer:
m
a table needs to be filled in order to obtain several values for
the analysis. In order to obtain regression coefficient estimates

0
,
1
,
2
, and
3
, it is necessary to solve the given
x
Where;

kj
u
j
j 1
+
k
k = 1to i (8)
simultaneous system of linear equations.
n
0
+
1
X
1i
+
2
X
2i
+
3
X
3i
Y
i

0
X
1i
+
1
X
2
+ X X +
X X X Y
2
(1)
(2)
m = number of input nodes
n = number of hidden nodes
i = number of output nodes
u = input node values
v = hidden node values

0
X
2i
+
1
X
1i
X
2i
+
2
X
2i
+
3
X
2i
X
3i
X
2i
Y
i
(3)
2
= synaptic weight
= threshold

0
X
3i
+
1
X
1i
X
3i
+
2
X
2i
X
3i
+
3
X
3i

X
3i
Y
i
(4)
The simultaneous system of linear equations above can be
simplified into matrix form. The values of regression
coefficients estimated can be obtained easier then.
In back-propagation neural network, the learning algorithm
has two phases. First, a training input pattern is presented to
the network input layer. The network then propagates the
input pattern from layer to layer until the output pattern is

n X
1i
X
2i
X
3i
1

Y
i

generated by the output layer. If this pattern is different from

X
2

X X

X X
1

the desired output, an error is calculated and then propagated

1i
1i 1i 2i
2
1i 3i
1
'
1
;

'
X
1i
Y
i
;
backwards through the network from the output layer to the

X
2i
X
1i
X
2i
X
2i
X
2i
X
3i 1

X
2i
Y
i

X
3i

X
1i
X
3i

X
2i
X
3i
2
1


3i ]

X
3i
Y
i
input layer. The weights are modified as the error is
propagated.
After the simultaneous system of linear equations above is
solved the regression coefficient estimates will be substitute to
the following regression model for surface roughness.
As with any other neural network, a back-propagation one
is determined by the connections between the neuron (the
networks architecture), the activation function used by the
neurons, and the learning algorithm (or the learning law) that
specifies the procedures for adjusting weights.
Y
i

0
+
1
X
1i
+
2
X
2i
+

3
X
3i
Where;
Y
i
= Surface Roughness (m)
X
1i
= Spindle Speed (rpm)
X
2i
= Feed Rate (mm/min)
X
3i
= Depth of Cut (mm)
(5)
Typically, a back-propagation network is multilayer
network that has three or four layers. The layers are fully
connected, that is, every neuron in each layer is connected to
every other neuron in the adjacent forward layer. Figure 3.2
shows the neural network computational model. The neural
network computational model coding is built using MATLAB
2008 software.
When the mathematical model is obtained, the value of
predicted surface roughness for each experiments can be
calculated.
B. Artificial Neural Network (ANN)
Artificial Neural Network is an adaptable system that can
learn relationships through repeated presentation of data and is
capable of generalizing to new, previously unseen data. Some
Spindle
Speed
Feed
Rate
Depth of
Cut
Input
layer
Hidden
layer
Output
Surface
Roughness
networks are supervised, in that a human determines what the
network should learn from the data.
For this study, the network is given a set of
inputs and corresponding desired outputs, and the
network tries to learn the input-output relationship by adapting
its free parameters.
The activation function f(x) used here is the sigmoid
function which is given by:
1
f ( x)
1 + exp(x)
Between the input and hidden layer:
Figure 1: Neural network computational model
1i
2i
3i
C. Experimental Design
To achieve the project objectives, multiple regression
analysis is used for statistical method and Artificial Neural
Network is used as artificial intelligent method. The
experiment is performs by using a FANUC CNC Milling -
14E. The workpiece tested is 6061 Aluminum
400mmx100mmx50mm. The end-milling and four flute high
speed steel is chooses as the machining operation and cutting
tool. The diameter of the tool is D=10mm. Three levels for
each variable are used. For spindle speed 1000, 1250 and 1500
rpm, for feed rate 152, 380 and 588 mm/min, and for depth of
cut 0.25, 0.76 and 1.27 mm.
IV. RESULTS
Surface roughness for each experiment is measured using
portable surface finish tester, Perthometer S2. All the data
taken from the experiment are shown in Table 2.
A. Multiple Regression Analysis Model
Since the surface roughness from the experiment has
been established, the analysis for multiple regression using
Equation xx above is done to obtain regression coefficient
0
,

1
,
2
, and
3
. The sum values calculated for X
1
,
X
1i
= 33750
X
2i
= 10080
X
3i
= 20.52
Y
i
= 58.751
X
2
= 43312500
X
2
= 4619232
X
2
= 20.277
X
1i
X
2i
= 12600000
X
1i
X
3i
= 25650
X
2i
X
3i
= 7660.8
X
1i
Y
i
= 72226.5
X
2i
Y
i
= 25350.268
X
3i
Y
i
= 44.19635
Substituting all the sums values into the simultaneous equation
of linear system (Equation 1-4)
( 27 )
0
+
1
(33750 ) +
2
(10080 ) +
3
( 20 .52 ) 58 .751

0
(33750) +
1
(43312500) +
2
(12600000) +
3
(25650)
72226.5

0
(10080) +
1
(12600000) +
2
(4619232) +
3
(7660.8)
25350.268

0
(20.52) +
1
(25650) +
2
(7660.8) +
3
(20.277) 44.19635
Transform above equations into matrix form;
Figure 2: FANUC CNC Milling -14E Workstation
27 33750 10080 20.52 1
0

58.751
1

33750 43312500 12600000 25650

1

72226.5

For this research, Full Factorial Experiment (FUFE) is

1008
0
12600000 4619232 7660.8
1
'

;

' ;
25350.268

2

applied. Full Factorial Experiment is the experiment where all
25650 7660.8 20.277
1

44.19635
the possible combinations levels of factors are realized. The

table below is the Full Factorial Experiments table for this

] 3
research. The parameters considered are Spindle Speed, Feed
Rate, and Depth of Cut. Thus, the numbers of experiment need
to be executed are N = 3
k
= 3
3
= 27 experiments.
TABLE 1: THE LEVELS OF EACH PARAMETER
After completing the solution for the matrix form, the
regression coefficients estimated are;

0
= 2.1066

1
= -0.0011

2
= 0.0040
Independent
Variables
Levels
Low Medium High

3
= -0.0971
Spindle Speed (rpm) 1000 1250 1500
Feed Rate (mm/min) 152 380 588
Depth of Cut (mm) 0.25 0.76 1.27
Depth of
Cut
(mm)
Surface
Roughness
(m)
0.25 1.173
0.76 1.681
1.27 1.275
0.25 2.265
0.76 2.443
1.27 2.367
0.25 3.84
0.76 3.586
1.27 3.307
0.25 1.276
0.76 1.301
1.27 1.603
0.25 2.392
0.76 2.138
1.27 2.137
0.25 3.637
0.76 2.469
1.27 2.773
0.25 0.64
0.76 1.224
1.27 1.301
0.25 2.392
0.76 1.808
1.27 2.215
0.25 2.723
0.76 2.316
1.27 2.469
i
TABLE 2: SURFACE ROUGHNESS OBTAINED FROM THE
EXPERIMENTS
B. ANN Model
Besides using multiple regression in surface roughness
No of
Experiment
Spindle
Speed
(rpm)
Feed Rate
(mm/min)
prediction, artificial neural network a branch of artificial
intelligent has been implemented as an alternative approach.
The predicted surface roughness has been perform using
1 1000 152
2 1000 152
3 1000 152
4 1000 380
5 1000 380
6 1000 380
7 1000 588
8 1000 588
9 1000 588
10 1250 152
11 1250 152
12 1250 152
13 1250 380
14 1250 380
15 1250 380
16 1250 588
17 1250 588
18 1250 588
19 1500 152
20 1500 152
21 1500 152
22 1500 380
23 1500 380
24 1500 380
25 1500 588
26 1500 588
27 1500 588
artificial neural network code in MATLAB 2008. Table 4.6
The input data for three independent variables spindle
speed, feed rate, and depth of cut while actual surface
roughness acted as target. The network propagates the input
pattern from layer to layer until the output is generated. Then
the result output will be compared with the target which is
and propagated back through network. Then, the weight will
be changed and the same process repeated until the smallest
error is achieved.
The plot of predicted surface roughness (output) against the
actual surface roughness (target) in Figure 3 below shown that
roughness is approaching towards the actual surface roughness
with the coefficient of determination, R is 0.98508.
Then, the regression coefficient can be substituted into the
general equation for multiple regression which shown as
equation 3.5 in previous chapter. The mathematical model
obtains to predict surface roughness is;
= 2.1066 0.0011X
1
+ 0.0040X
2
0.00971X
3
(9)
Where;
= Surface Roughness (m)
X
1
= Spindle Speed (rpm)
X
2
= Feed Rate (mm/min)
Figure 3: Predicted surface roughness against the actual
surface roughness
Table 3 shows comparison between actual and predicted
surface roughness using multiple regression analysis (MRA)
and artificial neural network (ANN). To measure the accuracy
for both prediction models, average error for both models is
calculated as follows.

X
3
= Depth of Cut (mm)

Ra
i

Ra
i

100% (10)
i
Ra
1 1.173 1.590325 1.1309
2 1.681 1.540804 1.518
3 1.275 1.491283 1.5166
4 2.265 2.502325 2.3069
5 2.443 2.452804 2.6525
6 2.367 2.403283 2.3639
7 3.84 3.334325 3.8318
8 3.586 3.284804 3.645
9 3.307 3.235283 3.2751
10 1.276 1.315325 1.2533
11 1.301 1.265804 1.3074
12 1.603 1.216283 1.6102
13 2.392 2.227325 2.4619
14 2.138 2.177804 2.3222
15 2.137 2.128283 2.0358
16 3.637 3.059325 3.5351
17 2.469 3.009804 2.5429
18 2.773 2.960283 2.6877
19 0.64 1.040325 0.824
20 1.224 0.990804 0.9998
21 1.301 0.941283 1.1844
22 2.392 1.952325 2.3616
23 1.808 1.902804 1.9984
24 2.215 1.853283 1.996
25 2.723 2.784325 2.5754
26 2.316 2.734804 2.6396
27 2.469 2.685283 2.5402
Where;

i
= Percentage error for each
experiment
Ra
i
= Experimental surface roughness
From the average percentage error calculated, the
effectiveness of each method can be determined and can be
compared.
m

Ra
i
= Predicted surface roughness
TABLE 3: COMPARISON BETWEEN ACTUAL AND PREDICTED
SURFACE ROUGHNESS

i
1
i
m
Where;
(11)
No of
Experiment
Actual
Surface
Roughness
Predicted Surface
Roughness (m)
= Average percentage error
m = number of experiments
( m)
MRA ANN
For this problem, average percentage error for MRA model is
13.3%, while for ANN model, the average percentage error is
6.42%.
Figure 4 shows the plot of predicted surface roughness
using multiple regression analysis and neural network, and
actual surface roughness. It shows that the value predicted
using neural network predict closely with the actual surface
roughness obtained from experiment. Nevertheless, the
prediction using multiple regression is also reliable and can be
accepted as a successive method.
V. CONCLUSION
The main purpose of this research is to provide an effective
and accurate way to predict surface roughness in CNC end
milling.
Multiple regression analysis had been applied to develop a
mathematical model for surface roughness prediction method.
The result of average percentage error is 13.3%, showing that
the prediction accuracy is about 86.7%. That means the model
developed is reliable to predict surface roughness with
accepting accuracy range based on previous research.
Artificial neural network is a feasible technique and has
been used quite often in recent researches in engineering field.
The adaptation of this technique provides a brand new
perspective in this field and in surface roughness prediction to
be precise. Back-propagation neural network had been
implemented to achieve the goal which is to minimize the
error during surface roughness prediction. The result of the
prediction is favorable with 6.42% average percentage of
error, meaning that neural network is capable to predict the
surface roughness up to 93.58% accurate. As a conclusion,
artificial neural network provided better accuracy to predict
surface roughness in CNC milling process.
Figure 4: Plot of predicted using multiple
regression and neural network and actual
experimental surface roughness